PerturbAI and OpenAI move into joint drug-target discovery after AI agent trial

PerturbAI, a San Francisco biotech that describes itself as AI-native, said on 21 October 2026 that its year-long collaboration with OpenAI will move into joint drug-target discovery, after the two companies published results from a system of AI reasoning agents built with Scripps Research¹.

What the collaboration produced

The partners built thousands of agents that each analyse cellular responses, search the scientific literature and link experimental results to existing knowledge, then combine the findings into what they call a reasoning atlas connecting genes, cell responses, prior studies and possible mechanisms¹. Run on PerturbAI's whole-brain genomic dataset, the system finished its analysis in 32 minutes, against an estimated months to a year of serial human analysis and literature review, by the companies' own estimate¹. OpenAI's head of life sciences, Joy Jiao, said the work involved "nearly 2,000 agents in parallel"¹.

Human biology experts rated the agents' reports for novelty, evidential support, follow-up priority and confidence. In blinded comparisons with reports produced by single AI agents, the experts preferred the multi-agent interpretations 75% of the time¹. The results appear in a joint preprint and blog post¹.

The dataset behind the work is PerturbAI's 8-million-cell, brain-wide CRISPR atlas, launched in March 2026, which the company says has been downloaded more than 40,000 times¹. PerturbAI has since added proprietary atlases in metabolic diseases and aging, and says it is already pursuing several targets surfaced by its engine¹.

What it means for the market

The deal pairs a frontier model developer with a company whose asset is proprietary causal data generated in living tissue. For OpenAI, it is a route into target discovery that depends on a partner's data rather than its own laboratories. For PerturbAI, it is a validation of the data-plus-model approach from the best-known AI vendor, at a very early stage of the company's life: it closed an oversubscribed pre-seed round in Q3 2025¹.

For investors and pharma business development teams, the pattern matters more than the single result. The bottleneck PerturbAI's chief executive, Grace Zheng, points to is interpretation rather than data generation: experiments now produce data "faster than scientists can interpret it"¹. If multi-agent analysis holds up, the value in AI drug discovery shifts further towards whoever owns large, causal, proprietary datasets.

The release gives no financial terms for the expanded collaboration, no named targets and no therapeutic programmes, so its commercial scale is unknown¹. The 32-minute and 75% figures come from the companies themselves, and the findings are in a preprint rather than a peer-reviewed journal¹. The expert comparison was against single AI agents, not against human analysts. The next step the companies name is applying the agents to PerturbAI's causal in vivo data to find new drug targets¹.